Existing research confirms the vulnerability of deep object detection models to gradient-based adversarial patch attacks, which manipulate detection outputs through carefully crafted perturbations. This process becomes prohibitively time-consuming, especially when multiple models are integrated for adversarial training. We propose a Variable Step-size Method (VSM) that dynamically adjusts attack step sizes based on adversarial loss ratios between perturbed and clean samples. Integrating VSM with PGD and MI-FGSM frameworks enables efficient generation of adversarial patches while maintaining attack transferability. Benchmark evaluations demonstrate VSM achieves 12–15% faster convergence than state-of-the-art methods while preserving 98–102% of baseline attack success rates across mainstream detection architectures. The adaptive step-size mechanism provides superior computational efficiency over conventional fixed-step approaches without compromising attack effectiveness.